A Drone Traffic Data Acquisition Method Based on Dynamic Privacy Protection and Ground-Air Collaboration

Abstract

This invention relates to the field of intelligent transportation data acquisition and privacy protection technology, and proposes a UAV traffic data acquisition method and system based on dynamic privacy protection and ground-air collaboration. It effectively solves the problems of high privacy leakage risk, incomplete coverage of complex scenarios, and weak cross-regional model generalization ability in traditional traffic data acquisition, and is applicable to diverse scenarios such as traffic management, logistics optimization, and academic research.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation data acquisition and privacy protection technology, specifically to a method and system for UAV traffic data acquisition that integrates dynamic privacy masking, data localization federated learning, and ground-air collaborative virtual-real dual-domain approach. It is applicable to high-security, high-coverage, and high-precision data acquisition needs in scenarios such as traffic management, logistics optimization, and academic research. Background Technology

[0002] With the widespread application of drones in traffic monitoring, traditional data collection methods face the following technical bottlenecks:

[0003] Privacy risk: Videos collected by drones contain sensitive information such as faces and license plates. Directly uploading or storing the raw data can easily lead to the leakage of personal information.

[0004] Incomplete scene coverage: A single drone's perspective has blind spots (such as no-fly zones or obstructed areas) and is difficult to cover complex environments (such as nighttime, rain, or fog), resulting in missing data.

[0005] Low data utilization efficiency: The traditional "centralized cloud training" model requires the aggregation of raw data from multiple regions, which has problems such as high transmission costs and high risk of privacy leakage, and the cross-regional model generalization ability is insufficient.

[0006] While existing technologies (such as YOLO object detection and static blurring desensitization) can partially solve the detection problem, they cannot simultaneously meet the multi-objective requirements of "privacy protection, data availability, and scene coverage," and a comprehensive solution is urgently needed. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a method and system for UAV traffic data acquisition based on dynamic privacy protection and ground-air collaboration. Through a technology chain of "dynamic privacy mask dual-track acquisition - data localization federated learning - ground-air collaborative virtual-real dual-domain expansion," it solves the problems of privacy leakage, incomplete scene coverage, and weak cross-regional model generalization ability in existing technologies. By integrating these technologies, a multi-objective balance of "privacy protection, data availability, and scene coverage" is achieved. In practical applications, it can support real-time traffic situation awareness, logistics route optimization, and accident emergency response, providing a new technical solution with high security, high coverage, and high generalization for the refined management of intelligent transportation. Specifically, this invention adopts the following technical solution:

[0008] A method for collecting traffic data by unmanned aerial vehicles (UAVs) based on dynamic privacy protection and air-ground collaboration includes the following steps:

[0009] (1) Dynamic privacy mask and traffic key data dual-track collection: The lightweight target detection model identifies sensitive areas (faces, license plates) and ordinary targets (vehicles, pedestrians) in the drone video in real time. Based on environmental parameters and business needs, the fuzzing parameters are dynamically adjusted to adaptively fuzz the sensitive areas, while extracting traffic key data (trajectory, vehicle type, lane occupancy) of ordinary targets.

[0010] (2) Data localization and privacy-enhanced federated learning: Edge nodes in each region train traffic prediction models based on local desensitized data, and only upload encrypted model parameters to the cloud. A cross-regional global model is generated through federated aggregation and differential privacy technology.

[0011] (3) Ground-to-air collaboration and virtual-real dual-domain data acquisition: integrate UAV video, roadside sensor and vehicle terminal data, complete privacy protection fusion through edge multi-party secure computing (MPC), and expand missing scenario data such as no-fly zones by combining digital twin and dynamic adversarial generation network (DAGN).

[0012] Preferably, step (1) of "dual-track collection of dynamic privacy mask and key traffic data" specifically includes:

[0013] (1.1) Use the improved YOLOv8-MobilenetV4 lightweight model to perform target detection on UAV video frames and output the target category (face, license plate, vehicle, pedestrian) and bounding box coordinates;

[0014] (1.2) Obtain environmental parameters (light intensity L, rain / fog level R) and business requirement parameters (e.g., S=1 around the school), and calculate the fuzzy parameters using the following formula:

[0015] k=k0·α·β·γ

[0016] σ=σ0·α·β·γ

[0017] Wherein, k0 and σ0 are the base kernel size and standard deviation, α is the class weight (face 1.2, license plate 1.0), β is the environmental correction factor β=1-0.1R+0.05L, and γ is the business demand factor (school periphery 1.5, ordinary area 1.0); (1.3) Gaussian blur or mosaic processing is applied to sensitive areas (face, license plate), and key data such as trajectory, vehicle type, and lane occupancy are extracted from ordinary target areas through optical flow method and high-precision map matching, and stored in structured format D={T,C m O l}

[0018] Preferably, step (2) of "data localization + privacy-enhanced federated learning" specifically includes:

[0019] (2.1) Edge nodes in each region are based on local de-identified data Di Training LSTM prediction model M i The loss function is:

[0020]

[0021] Among them, P i,j This is the predicted probability value for congestion. For real-world congestion labels, R i,j This is the predicted value of the accident risk. The label represents a real accident, and λ is the balance coefficient.

[0022] (2.2) Edge nodes use homomorphic encryption (such as the Paillier algorithm) to process the model parameters θ. i Encryption generation θ i' =Enc(θ) i Upload to the cloud;

[0023] (2.3) The cloud aggregates parameters using the FedAvg algorithm to generate a global model θ. global The formula is: Where, N i Let N be the amount of data in region i. total =∑N i ;

[0024] (2.4) Before uploading the parameters, perform gradient g i Add Laplace noise g i The formula is:

[0025] Where △g is the gradient sensitivity, and ∈ is the dynamically adjusted privacy budget.

[0026] Preferably, step (3) of "ground-air collaboration + virtual-real dual-domain data acquisition" specifically includes:

[0027] (3.1) Construct a ground-to-air sensing network consisting of "UAV + roadside sensors + vehicle-mounted terminals" to collect high-altitude video V u Radar point cloud T r Vehicle-mounted emergency braking incident B o ;

[0028] (3.2) Encrypting V at the edge node via the GMW protocol u T r B o Perform privacy-preserving fusion and output the incident correlation feature F = v min ,d min B count (Minimum relative speed, minimum distance between vehicles, number of emergency braking attempts);

[0029] (3.3) Construct a digital twin system for urban transportation based on OpenStreetMap, and generate virtual no-fly zone data V through Dynamic Generative Adversarial Network (DAGN). sim The adversarial training loss function between generator G and discriminator D is:

[0030]

[0031] Where x represents real data and z represents random noise. The generated virtual data is mixed with real data at a ratio of ρ = 0.3 to expand the training set D. total =D real ∪ρ·D sim .

[0032] A drone traffic data acquisition system for implementing the method includes: a data acquisition module consisting of a drone, roadside sensors, and a vehicle-mounted terminal, used to acquire high-altitude video, radar point clouds, and vehicle event data; a privacy processing module integrating a YOLOv8-MobilenetV4 model and a dynamic fuzzy parameter calculation unit, used for sensitive area identification, fuzzy processing, and key data extraction; and a federated learning module consisting of edge nodes (local model training unit and parameter encryption unit) and a cloud (parameter aggregation unit and differential privacy enhancement unit), used for cross-regional model collaborative training.

[0033] Virtual-Real Expansion Module: Includes a digital twin system and a dynamic adversarial generative network (DAGN), used for data generation and fusion in missing scenarios such as no-fly zones.

[0034] Preferably, the privacy processing module includes: a target detection submodule: using a lightweight YOLOv8-MobilenetV4 model to output target category and bounding box coordinates; a fuzzy parameter calculation submodule: calculating fuzzy parameters based on environmental parameters L, R and business requirements S using the formulas k = k0·α·β·γ and σ = σ0·α·β·γ; and a key data extraction submodule: extracting structured data such as trajectory, vehicle type, and lane occupancy through optical flow and high-precision map matching.

[0035] Preferably, the federated learning module includes: an edge training unit that trains an LSTM model based on local de-identified data and supports uploading homomorphic encrypted parameters; a cloud aggregation unit that aggregates encrypted parameters using the FedAvg algorithm and generates a global model using differential privacy technology; and a model distribution unit that distributes the global model to each edge node and supports dynamic model updates.

[0036] Preferably, the virtual-real expansion module includes: a digital twin submodule: constructing a virtual urban traffic scene based on OpenStreetMap; a DAGN generation submodule: generating virtual no-fly zone data through adversarial training of the generator (G) and discriminator (D); and a data fusion submodule: mixing virtual data with real data at a ratio of ρ = 0.3 to expand the training set.

[0037] The present invention has the following beneficial effects:

[0038] (1) The dynamic privacy mask and key data dual-track acquisition technology of the present invention identifies sensitive information (such as face and license plate) and key traffic data (such as vehicle trajectory and lane occupancy) in real time through a lightweight target detection model. It dynamically adjusts the fuzzy strategy in combination with the environment and business needs, so as to retain the data analysis value while strictly protecting privacy.

[0039] (2) Data localization federated learning mechanism: each region only uploads encrypted model parameters to the cloud for collaborative training, avoiding leakage of raw data and significantly improving the accuracy of cross-regional traffic prediction.

[0040] (3) Ground-air collaboration and virtual-real dual-domain data collection, integrating data from drones, roadside sensors and vehicle terminals, and combining digital twin technology to generate virtual data for missing scenarios such as no-fly zones, thus filling the coverage blind spots of traditional single devices. Attached Figure Description

[0041] Figure 1 System architecture diagram.

[0042] Figure 2 Dynamic fuzzy flowchart.

[0043] Figure 3 Federated learning structure diagram. Detailed Implementation

[0044] The technical solution of the present invention includes the following core steps (such as...) Figure 1 As shown):

[0045] Step 1: Dual-track collection of dynamic privacy mask and key traffic data

[0046] Objective: To protect sensitive information (faces, license plates) while retaining key data required for traffic analysis (such as vehicle trajectories and lane occupancy).

[0047] 1.1 Multi-target detection and sensitive area localization:

[0048] An improved lightweight YOLOv8-MobilenetV4 model is used to perform real-time detection of drone video frames, outputting target categories (pedestrians, vehicles, license plates, faces) and bounding box coordinates. The model structure is as follows:

[0049] YOLOv8-MobilenetV4(x)=Backbone(x)+Neck(x)+Head(x)

[0050] Backbone uses depthwise separable convolutions from MobileNetV4 to reduce computation, Neck is the SPPF (Spatial Pyramid Pooling-Fast) multi-scale feature fusion module, and Head is the decoupled detection head, which outputs the target category confidence C, bounding box coordinates (x1, y1, x2, y2), and key point coordinates (such as facial features).

[0051] 1.2 Dynamic Privacy Masking Strategy:

[0052] Based on the target category (face / license plate), environmental parameters (light intensity L, rain / fog level R), and business requirements (e.g., S=1 around a school), the fuzzing parameters (Gaussian kernel size k, standard deviation σ) are dynamically adjusted. Figure 2 As shown. The formula for calculating fuzzy intensity is:

[0053] k=k0·α·β·γ, (1)

[0054] σ=σ0·α·β·γ, (2)

[0055] Where: γ and σ0 are the base kernel size and standard deviation (e.g., k0 = 51, σ0 = 1.5); α is the class weight (α = 1.2 for faces, α = 1.0 for license plates); β is the environmental correction factor (β = 1 - 0.1R + 0.05L, R ∈ [0,3], L ∈ [0,1]);

[0056] γ is the business demand factor (γ = 1.5 for areas surrounding schools, γ = 1.0 for general areas).

[0057] 1.3 Extraction of Key Traffic Data:

[0058] For non-sensitive regions (such as vehicle outlines and pedestrian trajectories), the displacement Δ(x,y) between consecutive frames is calculated using the optical flow method Lucas-Kanade(I1,I2). Combined with the target detection results, the trajectory is extracted. Model C m Lane occupancy O l Key data such as (matching with high-precision maps) are stored in a structured format D = {T, C}. m O l}

[0059] Step 2: Data Localization + Privacy-Enhanced Federated Learning

[0060] Objective: To improve the accuracy of cross-regional traffic prediction by collaborating multi-regional models without transmitting the original data.

[0061] 2.1 Local model training:

[0062] Each region's edge nodes (such as roadside units) are based on locally anonymized data D i (Includes only key data T and C) m O l ), train LSTM prediction model M i Output the probability P of congestion in the next 15 minutes. i,j Or accident risk R i,j The model loss function is:

[0063]

[0064] in, The true label is λ, which is the balance coefficient λ = 0.3.

[0065] 2.2 Encrypted Parameter Upload and Aggregation:

[0066] Each edge node uses homomorphic encryption (such as the Paillier algorithm) to encrypt the model parameters θ i Encrypt, generate θ i' =Enc(θ) i The data is uploaded to the cloud. The cloud uses a federated averaging (FedAvg) algorithm to aggregate parameters, such as... Figure 3 As shown:

[0067]

[0068] Where, N i Let N be the amount of data in region i. total =∑N i .

[0069] 2.3 Differential Privacy Enhancement:

[0070] Before uploading the parameters, the gradient g is... i Add Laplace noise ∈ (privacy budget) to ensure:

[0071]

[0072] Where △g is the gradient sensitivity △g=1, ∈ dynamically adjusted (e.g., during busy periods ∈=0.5, during idle periods ∈=1.0).

[0073] Step 3: Ground-to-air collaboration + virtual-real dual-domain data acquisition mechanism

[0074] Objective: To supplement data on blind spots such as no-fly zones and improve scene coverage capabilities.

[0075] 3.1 Collaborative acquisition of multi-source ground and air data:

[0076] Constructing a ground-to-air perception network consisting of "UAVs (high-altitude video) + roadside sensors (radar point clouds) + on-board units (OBUs)": UAVs: collecting high-altitude video V u Output key desensitization data D u Roadside radar: outputs vehicle trajectory T r Speed ​​v r Onboard OBU: Outputs emergency braking event B o Steering angle θ o .

[0077] 3.2 Edge-based Multi-Party Secure Computation (MPC) Fusion:

[0078] Privacy-preserving fusion of multi-source data is performed at edge nodes (such as roadside units) using the GMW protocol: Input: Encrypted D u’ T r’ B o Output: F = v, the fused accident correlation features min ,d min B count (Minimum relative speed, minimum distance between vehicles, number of emergency braking operations).

[0079] 3.3 Expansion of Virtual and Real-Domain Scenarios:

[0080] A digital twin system for urban transportation is built based on OpenStreetMap, and virtual no-fly zone data (V) is generated through Dynamic Generative Adversarial Network (DAGN). sim In DAGN, the generator G and discriminator D are trained adversarially, with the loss function being:

[0081]

[0082] Where x represents real data and z represents random noise. The generated virtual data is mixed with real data at a ratio of ρ = 0.3 to expand the training set D. total =D real ∪ρ·D sim .

[0083] Compared with the prior art, the present invention has the following advantages:

[0084] Balancing privacy protection and data availability: Dynamic privacy masking technology, through adaptive fuzzy parameter adjustment (Formula 1-2), achieves a face / license plate anonymization rate of ≥99% while retaining key data (such as trajectory continuity rate and vehicle model recognition accuracy) of ≥95%, meeting the requirements of the Personal Information Protection Law.

[0085] Improved generalization ability of cross-regional models: Data-local federated learning (Formula 4-5) avoids uploading raw data to the cloud, and the accuracy of cross-regional congestion prediction is improved by 15% compared with the centralized training model (mAP increases from 78% to 93%), with a privacy attack success rate of ≤5%.

[0086] Enhanced scene coverage: The ground-air collaboration + virtual-real dual-domain mechanism (Formula 8-9) enables the scene completion rate of no-fly zones to be ≥90%, the data missing rate in complex environments (rain, fog, night) to be reduced from 25% to 5%, and the accident early warning accuracy rate to be ≥85%.

[0087] The implementation process of this invention is illustrated below using a specific scenario (traffic monitoring in western Beijing):

[0088] Data collection phase:

[0089] Deploy 10 drones (as in step 3.1) to cover areas such as Haidian and Shijingshan. Use the YOLOv8-MobilenetV4 model to detect pedestrians and vehicles in real time (step 1.1). Apply dynamic blurring to faces / license plates (formulas 1-2) to extract key data such as trajectories and vehicle types (step 1.3).

[0090] Data fusion and model training:

[0091] The roadside unit fuses data from drones, radar, and OBU via the MPC protocol (step 3.2) to generate accident-related features; edge nodes in each area train LSTM models based on local data (step 2.1), encrypt and upload parameters to the cloud for aggregation (step 2.2), and combine differential privacy enhancement (step 2.3).

[0092] Scenario Expansion and Validation:

[0093] A digital twin system was built based on OpenStreetMap. Virtual data of the no-fly zone was generated by DAGN (step 3.3). The data was mixed with real data to train the model and verify the no-fly zone scene completion rate and accident warning accuracy.

Claims

1. A method for collecting traffic data from unmanned aerial vehicles (UAVs) based on dynamic privacy protection and air-ground collaboration, characterized in that, Includes the following steps: (1) Dynamic privacy mask and traffic key data dual-track collection: The sensitive areas and ordinary targets in the drone video are identified in real time through a lightweight target detection model. The fuzzing parameters are dynamically adjusted based on environmental parameters and business needs. The sensitive areas are adaptively fuzzed, while the traffic key data of ordinary targets are extracted. (2) Data localization and privacy-enhanced federated learning: The edge nodes of each region train the traffic prediction model based on local desensitized data. Only the encrypted model parameters are uploaded to the cloud. A cross-regional global model is generated through federated aggregation and differential privacy technology. (3) Ground-air collaboration and virtual-real dual-domain data collection: The drone video, roadside sensor and vehicle terminal data are integrated. Privacy protection fusion is completed through multi-party secure computation (MPC) at the edge. The missing scenario data of the no-fly zone is expanded by combining digital twin and dynamic adversarial generation network (DAGN). The data localization and privacy-enhanced federated learning mentioned in step (2) specifically includes: (2.1) Each regional edge node based on local de-identified data Training an LSTM prediction model The loss function is: in, This is the predicted probability value for congestion. Labels for actual traffic congestion. This is the predicted value of the accident risk. Labeled as a real accident. N is the balance coefficient, and N is the amount of data in the region; (2.2) Edge nodes use homomorphic encryption to adjust model parameters. Encryption generation ENC is uploaded to the cloud; (2.3) The cloud aggregates parameters using the FedAvg algorithm to generate a global model. The formula is: ,in, Let i be the amount of data in region i. (2.4) Gradients are processed before parameter upload. Adding Laplace noise The formula is: ,in, For gradient sensitivity, A dynamically adjusted privacy budget; The ground-air coordination and virtual-real dual-domain data acquisition mentioned in step (3) specifically includes: (3.1) Construct a ground-to-air sensing network consisting of "UAV + roadside sensors + vehicle-mounted terminals" to collect high-altitude video. Radar point cloud Vehicle emergency braking incident (3.2) Encryption at the edge node via GMW protocol , , Perform privacy protection integration and output accident correlation features. ;in v min 、d min 、B count These are the minimum relative speed, minimum vehicle distance, and number of emergency braking operations, respectively; (3.3) Based on OpenStreetMap, a digital twin system for urban traffic is constructed, and virtual data of the no-fly zone is generated through Dynamic Generative Adversarial Network (DAGN). The adversarial training loss function between generator G and discriminator D is: , Where x represents real data, z represents random noise, and the generated virtual data is proportional to the real data. =0.3 Mixing, expanding the training set .

2. The method according to claim 1, characterized in that, The dual-track collection of dynamic privacy mask and key traffic data in step (1) specifically includes: (1.1) using an improved YOLOv8-MobilenetV4 lightweight model to perform target detection on UAV video frames and output the target category and bounding box coordinates; (1.2) obtaining environmental parameters and business requirement parameters, and calculating fuzzy parameters using the following formula: in, , Based on the core size and standard deviation, For category weights, Environmental correction factor = 1 - 0.1R + 0.05L, where L is the light intensity and R is the rain / fog level. For business demand factors; (1.3) Apply Gaussian blur or mosaic processing to sensitive areas, and extract trajectory T and vehicle type for ordinary target areas through optical flow method and high-precision map matching. Lane occupation Key data is stored in a structured format D = { } 3. A UAV traffic data acquisition system for implementing the method of claim 1 or 2, characterized in that, include: Data acquisition module: Composed of drones, roadside sensors, and vehicle-mounted terminals, used to acquire high-altitude video, radar point clouds, and vehicle event data; Privacy processing module: integrates the YOLOv8-MobilenetV4 model and dynamic fuzzy parameter calculation unit for sensitive area identification, fuzzing, and key data extraction; Federated learning module: consists of edge nodes and the cloud, used for cross-regional model collaborative training; Virtual-Real Expansion Module: Includes a digital twin system and a dynamic adversarial generative network (DAGN), used for the generation and fusion of missing scenario data in no-fly zones.

4. The system according to claim 3, characterized in that, The privacy processing module includes: a target detection submodule: employing a lightweight YOLOv8-MobilenetV4 model to output target categories and bounding box coordinates; and a fuzzy parameter calculation submodule: based on environmental parameters L and R and business requirements S, using formulas... and Calculate fuzzy parameters; Key data extraction submodule: Extract structured data of trajectory, vehicle type, and lane occupancy through optical flow method and high-precision map matching.

5. The system according to claim 3, characterized in that, The federated learning module includes: an edge training unit that trains an LSTM model based on local de-identified data and supports uploading homomorphic encrypted parameters; a cloud aggregation unit that aggregates encrypted parameters using the FedAvg algorithm and generates a global model using differential privacy technology; and a model distribution unit that distributes the global model to each edge node and supports dynamic model updates.

6. The system according to claim 3, characterized in that, The virtual-real augmentation module includes: a digital twin submodule: constructing a virtual urban traffic scene based on OpenStreetMap; a DAGN generation submodule: generating virtual no-fly zone data through adversarial training of the generator (G) and discriminator (D); and a data fusion submodule: proportionally integrating virtual data with real data. Mix and expand the training set.

Images